Research map / Adversarial robustness and security / Certified and empirical robustness
Rethinking Fast Adversarial Training: A Splitting Technique To Overcome Catastrophic Overfitting
Masoumeh Zareapoor, Pourya Shamsolmoali
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DR splitting (RAT) to MEP-PGI at eps=16/255 | MEP-PGI · CIFAR10 (ResNet20, RA-AA) | RA-AA (%): 19.31 -> 23.91 | Improved |
| add DR splitting (RAT) to MEP-PGI at eps=16/255 | MEP-PGI · CIFAR10 (ResNet20) | standard accuracy (%): 54.23 -> 69.86 | Improved |
| add DR splitting (RAT) to Fast-AT at eps=16/255 | Fast-AT · CIFAR10 (ResNet20, RA-AA) | RA-AA (%): 14.97 -> 19.64 | Improved |
| add DR splitting (RAT) to PGI-BP on Tiny-ImageNet | PGI-BP · Tiny-ImageNet (PreActResNet18) | RA-AA (%): 4.18 -> 6.25 | Improved |
| add DR splitting (RAT) to PGI-BP on ImageNet | PGI-BP · ImageNet (ResNet50) | RA-PGD10 (%): 2.81 -> 8.70 | Improved |
| replace multi-step PGD-AT with RAT-MEP (single-step) on CIFAR10 | RAT-MEP · CIFAR10 (ResNet20) | RA-PGD50 (%) and time: 34.21 -> 33.64 RA-PGD50; 6.3 -> 1.7 hours | Mixed |
| replace PGD-AT with RAT-MEP on CIFAR100 AutoAttack | RAT-MEP · CIFAR100 (ResNet20) | RA-AA (%): 12.54 -> 11.49 | Hurt |
| apply RAT to OAAT | OAAT · CIFAR10 (ResNet20) | RA-PGD-10 SQ (%): 21.43 -> 24.66 | Improved |
| compare RAT to Fast-GA at small eps=8/255 (WideResNet34-10) | RAT · CIFAR10 (eps=8/255) | AA (%): 46.37 -> 46.21 | No clear effect |
| use higher regularization strength lambda (0.9 or 1) (ablation) | RAT · CIFAR10 (PreActResNet18, validation 1024 images) | PGD-20 accuracy | Mixed |
| use lower lambda with large eps (ablation) | RAT · CIFAR10 (PreActResNet18) | PGD-20 accuracy | Hurt |
| vary lambda within 0.3 to 0.6 (ablation) | RAT · CIFAR10 (PreActResNet18) | PGD-20 accuracy | No clear effect |
Closest papers by meaning
- Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspective (ECCV 2024 · similarity 0.91)
- Taxonomy Driven Fast Adversarial Training (AAAI 2024 · similarity 0.89)
- Efficient local linearity regularization to overcome catastrophic overfitting (ICLR 2024 · similarity 0.89)
- SORA: Free Second-Order Attacks in Fast Adversarial Training (ICML 2026 · similarity 0.89)
- Understanding and Improving Fast Adversarial Training against $l_0$ Bounded Perturbations (NeurIPS 2025 · similarity 0.89)
- Layer-Aware Analysis of Catastrophic Overfitting: Revealing the Pseudo-Robust Shortcut Dependency (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
